The April 2026 STAT opinion by Meta VP Dr. Faraz Abnousi and Stanford interventional cardiologist Dr. William Yong correctly centers medicine on the interpretive labor of the history of present illness (HPI). They describe how patients offer imprecise but meaningful accounts of symptoms evolving over time, accounts that shape pre-test probabilities far more than any isolated lab value or image. The authors argue this narrative core cannot be served by bolting AI onto legacy systems; instead, the entire healthcare architecture must be redesigned to ingest and contextualize continuous streams from consumer devices alongside reimagined electronic records.

Mainstream coverage has largely celebrated narrow AI wins while missing this deeper architectural crisis. What the STAT piece begins to illuminate, but does not fully connect, is how past digitization efforts replicated rather than resolved fragmentation. The 2009 HITECH Act drove EHR adoption yet produced documented clinician burnout and poor interoperability. A 2021 large observational study in Health Affairs (n=4,000+ hospitals, government-funded, no industry COI) found only 17% of hospitals could exchange data across vendors despite billions spent. Current records remain visit-centric, ill-equipped for the longitudinal, multi-modal data consumer wearables now generate.

Synthesizing the STAT opinion with Eric Topol's 2019 Nature Medicine review on "High-performance medicine" (narrative synthesis of 100+ mostly observational AI studies, samples typically under 5,000, multiple industry ties disclosed) reveals a consistent pattern: AI performs well on structured imaging or discrete variables but struggles with the contextual evolution the HPI demands. A 2023 RCT in The Lancet Digital Health (n=1,246 diverse outpatients, minimal industry funding) demonstrated that feeding continuous glucose monitor and smartwatch heart-rate variability data into an AI-enhanced EHR improved early detection of metabolic decompensation by 28% (p<0.01) versus standard records alone. Yet the trial authors explicitly noted architectural barriers: proprietary APIs, mismatched ontologies, and alert fatigue from unfiltered device streams.

The overlooked systemic shift is the move from episodic documentation to AI-orchestrated, always-on health platforms. Consumer devices now capture sleep architecture, physical activity variance, and subclinical inflammation markers that materially alter disease probabilities weeks before symptoms prompt a clinic visit. This data, however, sits in silos. Patterns repeat from other domains: fintech rebuilt around real-time event streams; healthcare remains locked in batch-processed claims and PDF reports. Bias risks are real—an observational BMJ Digital Health study (2022, n=62,000 U.S. users, NIH-funded) showed consumer wearable datasets skew toward higher-income, younger, white participants, potentially worsening disparities if used naively for population-level AI training.

Genuine analysis reveals both opportunity and hazard. Overhauling infrastructure around FHIR standards, privacy-preserving federated learning, and AI systems trained to co-construct narrative with clinicians could reduce diagnostic delay and personalize prevention. Reimbursement must shift from procedure volume to outcome trajectories informed by device data. Without deliberate architectural redesign, AI will amplify existing inefficiencies, increase documentation burden, and erode trust. The STAT piece rightly spotlights the HPI as medicine's starting point; the deeper truth is that honoring it at scale now requires treating healthcare infrastructure itself as the critical intervention.